Skip to contentSkip to conversation

Why idle_freq is set to RPn and not RPe

DISCUSSION Kamble, Sagar A 3 participants 9 messages Dec 30, 2015 – Jan 5, 2016

lore
Share & tools
lore
#0 lore raw

Hi Chris,

With below commit, idle frequency is made RPn (HW Min).
Why are we not keeping it at RPe (Efficient Frequency)?
My understanding was to set Rpe on idle so that when GPU is out of RC6 it can start operating at efficient frequency.

commit aed242ff7ebb697e4dff912bd4dc7ec7192f7581

Author: Chris Wilson <[email protected]>
Date:   Wed Mar 18 09:48:21 2015 +0000

drm/i915: Relax RPS contraints to allows setting minfreq on idle

Thanks
Sagar

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#1 lore raw
Quoted 6 lines · Sagar A

Hi Chris,

With below commit, idle frequency is made RPn (HW Min). Why are we not keeping it at RPe (Efficient Frequency)? My understanding was to set Rpe on idle so that when GPU is out of RC6 it can start operating at efficient frequency.

The driver is *idle*. When there is work to be submitted to the GPU, then we go back to RPe (though we wait for it to wake up first). RPe is just an inflexion point on the power curve.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

Hello Chris!

The question is: why this change in behavior was made?

On previous platforms Gfx Turbo frequency selection range in driver was in between Rpe & Rp0. Since Rpe is the possible Fmax at Vmin, it was used as the starting frequency once driver booted and any value lower than that was not requested.

Thanks!
Radek

Quoted 16 lines

-----Original Message-----
From: Chris Wilson [mailto:[email protected]]
Sent: Wednesday, December 30, 2015 10:31 AM
To: Kamble, Sagar A
Cc: S, Deepak; Szwichtenberg, Radoslaw; Intel Graphics Development; Goel, Akash
Subject: Re: Why idle_freq is set to RPn and not RPe

Quoted 6 lines · Sagar A

Hi Chris,

With below commit, idle frequency is made RPn (HW Min). Why are we not keeping it at RPe (Efficient Frequency)? My understanding was to set Rpe on idle so that when GPU is out of RC6 it can start operating at efficient frequency.

Quoted 7 lines

The driver is *idle*. When there is work to be submitted to the GPU, then we go back to RPe (though we wait for it to wake up first). RPe is just an inflexion point on the power curve.
-Chris

--
Chris Wilson, Intel Open Source Technology Centre

--------------------------------------------------------------------

Intel Technology Poland sp. z o.o.
ul. Slowackiego 173 | 80-298 Gdansk | Sad Rejonowy Gdansk Polnoc | VII Wydzial Gospodarczy Krajowego Rejestru Sadowego - KRS 101882 | NIP 957-07-52-316 | Kapital zakladowy 200.000 PLN.

Ta wiadomosc wraz z zalacznikami jest przeznaczona dla okreslonego adresata i moze zawierac informacje poufne. W razie przypadkowego otrzymania tej wiadomosci, prosimy o powiadomienie nadawcy oraz trwale jej usuniecie; jakiekolwiek przegladanie lub rozpowszechnianie jest zabronione.
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). If you are not the intended recipient, please contact the sender and delete all copies; any review or distribution by others is strictly prohibited.

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw
Quoted 3 lines · Radoslaw

Hello Chris!

The question is: why this change in behavior was made?

To take a defensive position to ensure minimum power consumption. The GPU is idle, the driver has no work for it to do, it should be power gated, but on the off-chance that fails pick the lowest power mode we have available for it to idle in.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw
In reply to Szwichtenberg, Radoslaw

Turbo frequency range is Rpe to Rp0 when GPU is active as, on workload submission frequency is taken to Rpe.

Does the HW require us to drop to RPn before entering RC6?
If we can enter RC6 even with other frequencies I think we can keep running at Rpe on Idle.

Only benefit of running at Rpn might be some additional power saving when workload is very lightweight. Is that correct Chris?

Thanks
Sagar

Quoted 31 lines · Radoslaw

Hello Chris!

The question is: why this change in behavior was made?

On previous platforms Gfx Turbo frequency selection range in driver was in between Rpe & Rp0. Since Rpe is the possible Fmax at Vmin, it was used as the starting frequency once driver booted and any value lower than that was not requested.

Thanks!
Radek

Quoted 22 lines

-----Original Message-----
From: Chris Wilson [mailto:[email protected]]
Sent: Wednesday, December 30, 2015 10:31 AM
To: Kamble, Sagar A
Cc: S, Deepak; Szwichtenberg, Radoslaw; Intel Graphics Development; Goel, Akash
Subject: Re: Why idle_freq is set to RPn and not RPe

Quoted 6 lines · Sagar A

Hi Chris,

With below commit, idle frequency is made RPn (HW Min). Why are we not keeping it at RPe (Efficient Frequency)? My understanding was to set Rpe on idle so that when GPU is out of RC6 it can start operating at efficient frequency.

The driver is *idle*. When there is work to be submitted to the GPU, then we go back to RPe (though we wait for it to wake up first). RPe is just an inflexion point on the power curve.
-Chris

--
Chris Wilson, Intel Open Source Technology Centre

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw
Quoted 6 lines · Sagar A

Turbo frequency range is Rpe to Rp0 when GPU is active as, on workload submission frequency is taken to Rpe.

Does the HW require us to drop to RPn before entering RC6? If we can enter RC6 even with other frequencies I think we can keep running at Rpe on Idle.

Remember that we quite frequently prevent the hardware going into RC6, and that it has been known for the hardware to fail to enter RC6 itself (through driver error or whatnot). Going to the extreme, why wouldn't you set Rp0 on idle, since that will give the best restart latency? -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#6 lore raw
Quoted 8 lines · Chris Wilson
Quoted 6 lines · Sagar A

Turbo frequency range is Rpe to Rp0 when GPU is active as, on workload submission frequency is taken to Rpe.

Does the HW require us to drop to RPn before entering RC6? If we can enter RC6 even with other frequencies I think we can keep running at Rpe on Idle.

Remember that we quite frequently prevent the hardware going into RC6,

I assume this is threshold times in TO/EI mode for which GT is idle but not power gated.

Quoted 2 lines

and that it has been known for the hardware to fail to enter RC6 itself (through driver error or whatnot).

And assume this is because of forcewake/rc6 setup errors in driver paths which should not happen in best case :) Agree that running at Rpn makes sense.

Quoted 2 lines

Going to the extreme, why wouldn't
you set Rp0 on idle, since that will give the best restart latency?

True. We can have different logic that starts from Rp0 and comes down if perf is met.

Quoted 2 lines

-Chris

Thanks for the inputs Chris.
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#7 lore raw

Hello Chris,

Happy New Year!
Thanks for answers so far. I have some additional questions.

You wrote that this change was made to take a defensive position to ensure minimum power consumption - did we do any power measurements to confirm the benefit? How does this change affect user experience and overall performance? Did we do any performance impact measurements? Is same change done in Windows implementation? Was this change in the algorithm discussed with HW team?

Thanks!
Radek

Quoted 26 lines

-----Original Message-----
From: Kamble, Sagar A
Sent: Thursday, December 31, 2015 6:38 PM
To: Chris Wilson; Szwichtenberg, Radoslaw; S, Deepak; Intel Graphics Development; Goel, Akash
Subject: Re: Why idle_freq is set to RPn and not RPe

Quoted 2 lines · Chris Wilson
Quoted 1 line · Sagar A

Turbo frequency range is Rpe to Rp0 when GPU is active as, on

workload

Quoted 6 lines
Quoted 5 lines

submission frequency is taken to Rpe.

Does the HW require us to drop to RPn before entering RC6? If we can enter RC6 even with other frequencies I think we can keep running at Rpe on Idle.

Remember that we quite frequently prevent the hardware going into RC6,

I assume this is threshold times in TO/EI mode for which GT is idle but not power gated.

Quoted 2 lines

and that it has been known for the hardware to fail to enter RC6 itself (through driver error or whatnot).

And assume this is because of forcewake/rc6 setup errors in driver paths which should not happen in best case :) Agree that running at Rpn makes sense.

Quoted 7 lines
Quoted 2 lines

Going to the extreme, why wouldn't
you set Rp0 on idle, since that will give the best restart latency?

True. We can have different logic that starts from Rp0 and comes down if perf is met.

Quoted 2 lines

-Chris

Thanks for the inputs Chris.

--------------------------------------------------------------------

Intel Technology Poland sp. z o.o.
ul. Slowackiego 173 | 80-298 Gdansk | Sad Rejonowy Gdansk Polnoc | VII Wydzial Gospodarczy Krajowego Rejestru Sadowego - KRS 101882 | NIP 957-07-52-316 | Kapital zakladowy 200.000 PLN.

Ta wiadomosc wraz z zalacznikami jest przeznaczona dla okreslonego adresata i moze zawierac informacje poufne. W razie przypadkowego otrzymania tej wiadomosci, prosimy o powiadomienie nadawcy oraz trwale jej usuniecie; jakiekolwiek przegladanie lub rozpowszechnianie jest zabronione.
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). If you are not the intended recipient, please contact the sender and delete all copies; any review or distribution by others is strictly prohibited.

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#8 lore raw
Quoted 6 lines · Radoslaw

Hello Chris,

Happy New Year!
Thanks for answers so far. I have some additional questions.

You wrote that this change was made to take a defensive position to ensure minimum power consumption - did we do any power measurements to confirm the benefit? How does this change affect user experience and overall performance? Did we do any performance impact measurements? Is same change done in Windows implementation? Was this change in the algorithm discussed with HW team?

None of the above. You can verify for yourself through code inspection that it will not impact ux.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx